suanmosuanyangtechnology/memorybear

MemoryBear is a Python-based system designed to give AI models a more persistent and adaptive form of memory. The core problem it addresses is that standard large language models forget earlier parts of a conversation once their context window fills up, cannot carry information from one session to the next, and share no memory across multiple agents working together on the same task. The system draws its design from biological cognition. It models memory encoding after the hippocampus, knowledge consolidation after the neocortex, and intentional forgetting after the synaptic pruning process the brain uses to discard redundant connections. The goal is not just to store information but to refine and connect it over time, turning raw conversation history into structured, queryable knowledge. The memory extraction engine parses unstructured text to pull out key facts as structured subject-action-object triples. Temporal anchoring attaches time markers to time-sensitive information so the system can track evolving facts. A graph-based storage layer connects related pieces of knowledge rather than treating each fact as isolated. An intelligent forgetting mechanism removes outdated or low-relevance data so the memory store does not grow stale. MemoryBear is built to work across multi-agent setups, where several AI agents collaborate on a task. Without shared memory, each agent would see only its own slice of context and could give contradictory or redundant responses. MemoryBear provides a shared memory layer that all agents can read from and write to, so users do not have to repeat the same information to each agent separately. The system is licensed under Apache 2.0 and requires Python 3.12 or later. A companion research paper describes the architecture in more detail. The README is written in English with a Chinese version linked separately.